Refiner workflow

[jwax33]

What is the workflow for using the SDXL Refiner in the new RC1.5? I don't see any option to enable it anywhere. Right now I'm sending base SDXL images to img2img, then switching to the SDXL Refiner model, and running that but results are not the greatest.

[howDoILearnPython]

i think the current RC did not support the use of refiner yet?

[dhwz]

It does somehow, but not the way it may be used by StabilityAI.

@jwax33 You've to lower the denoise strength to ~0,2-0,3 else you'll get a completely different image

[jwax33]

I've been using .3 which gives me pretty much the same image but the refiner has a really bad tendency to age a person by 20+ years from the original image. I've got a ~21yo guy who looks 45+ after going through the refiner.

The refiner does add overall detail to the image, though, and I like it when it's not aging people for some reason.

[dhwz]

Oh yeah I know what you mean, I'm not sure that is an issue with the process or it's just the refiner model itself.
Faces with base look completely different to me also.
Maybe compare results on ComfyUI with/without refiner model?

I personally prefer the base model results over the refiner.

[jwax33]

Installed ComfyUI and compared. Comfy's refiner process does very little people aging compared to the current img2img method in A1111.

[bjornlarssen]

I was using SDXL demo extension and noticed the same. It adds detail and 20 years.

Maybe I need to add "botox" to the prompt ;)

[binnn6]

@AUTOMATIC1111 maybe refiner should be processed in txt2img like hires.fix ?